Statistical Parity

Statistical parity is formally defined as requiring the probability of a positive outcome (e.g., receiving a loan, being hired) to be equal across groups defined by a protected attribute A. For a binary classifier with prediction Ŷ and binary protected attribute A (e.g., A=0 for group 1, A=1 for group 2), it is satisfied when:

P(Ŷ = 1 | A = 0) = P(Ŷ = 1 | A = 1)

This means the selection rate or acceptance rate is identical for all groups. It is a property of the model's predictions only, not the accuracy of those predictions.

At its core, statistical parity enforces statistical independence between the model's prediction (Ŷ) and the protected attribute (A). This is a strong condition that can be at odds with other objectives:

• Ignores Ground Truth: It does not consider whether the predictions are correct. A model can achieve perfect statistical parity while being completely inaccurate.
• Clashes with Calibration: A model is well-calibrated if, for individuals assigned a probability p, the proportion who truly belong to the positive class is p. Achieving both calibration and statistical parity is often mathematically impossible if the base rates (prevalence of the positive outcome) differ between groups.

Statistical parity is valued for its conceptual simplicity and enforceability, making it suitable for specific regulatory and high-stakes scenarios.

Key Strengths:

• Simple to Measure and Monitor: The metric is straightforward to compute from a model's output matrix.
• Legal and Regulatory Alignment: It directly corresponds to notions of disparate impact in anti-discrimination law (e.g., the 80% rule in U.S. employment law).
• Promotes Representation: In resource allocation (e.g., distributing benefits) or screening (e.g., resume shortlisting), it can ensure proportional representation from different demographic groups.

It is most appropriate when the goal is procedural fairness in access, rather than predictive accuracy based on historical outcomes.

Statistical parity is often criticized for being a blunt instrument that can enforce fairness at the cost of utility and individual justice.

Major Limitations:

• Ignores Legitimate Correlates: It can force a model to ignore proxy variables that are correlated with the protected attribute but are also legitimate predictors of the outcome (e.g., zip code correlating with both race and credit risk).
• Can Perpetuate Historical Bias: If historical data reflects societal bias (e.g., lower hiring rates for a group), a model achieving statistical parity may simply replicate that biased rate for all groups.
• Violates Individual Fairness: Two individuals with identical relevant qualifications may receive different predictions to meet the group quota, violating the principle that similar individuals should be treated similarly.
• Base Rate Problem: As mentioned, it conflicts with calibration when base rates differ, forcing a trade-off.

Achieving statistical parity typically requires explicit intervention during model training or post-processing.

Common Methods:

• Pre-processing: Modify the training data by reweighting, resampling, or transforming features to remove dependency on the protected attribute.
• In-processing: Add a fairness constraint or penalty term to the model's loss function that penalizes differences in group-wise prediction rates. Techniques include adversarial debiasing, where a second network tries to predict the protected attribute from the main model's predictions.
• Post-processing: Adjust the decision threshold for each protected group (e.g., using the Equalized Odds post-processing algorithm) until statistical parity is achieved. This is often the most straightforward but can be suboptimal.

Statistical parity is one point in a landscape of fairness definitions. Key alternatives address its shortcomings:

• Equalized Odds (Separation): Requires the model's true positive rate (TPR) and false positive rate (FPR) to be equal across groups. This considers the ground truth label (Y), allowing predictions to depend on A if justified by Y.
• Predictive Parity (Sufficiency): Requires precision (or positive predictive value) to be equal across groups. This ensures that the meaning of a positive prediction is consistent.
• Individual Fairness: A more granular concept requiring that similar individuals receive similar predictions, measured by a task-specific similarity metric.

These metrics often form impossibility triples—under realistic conditions, you cannot satisfy statistical parity, calibration, and equalized odds simultaneously.

Demographic parity is a direct synonym for statistical parity. It is a group fairness criterion requiring that the selection rate (probability of a positive prediction) is equal across groups defined by a protected attribute (e.g., race, gender).

• Key Principle: Independence between the prediction and the protected attribute.
• Example: If a loan approval model grants loans to 10% of applicants in Group A, it must also grant loans to 10% of applicants in Group B to satisfy demographic parity.
• Limitation: It does not consider potential differences in qualification rates between groups, which can lead to fairness through blindness that may not address underlying inequities.

Equal opportunity is a fairness metric that requires the true positive rate (recall) to be equal across protected groups. It focuses on ensuring qualified individuals have an equal chance of being correctly identified, regardless of group membership.

• Mathematical Definition: P(Ŷ=1 | Y=1, A=a) = P(Ŷ=1 | Y=1, A=b) for all groups a, b.
• Context: Used when the cost of a false negative (e.g., denying a qualified candidate a loan) is high and should be distributed equally.
• Contrast with Statistical Parity: Equal opportunity conditions on the actual positive outcome (Y=1), whereas statistical parity does not, making it suitable for scenarios where base rates differ between groups.

Equalized odds is a stricter fairness criterion than equal opportunity. It requires that both the true positive rate and the false positive rate are equal across protected attribute groups.

• Mathematical Definition: P(Ŷ=1 | Y=y, A=a) = P(Ŷ=1 | Y=y, A=b) for all groups a, b and for both y=0 and y=1.
• Implication: The model's errors (both false positives and false negatives) must be independent of group membership. This is a very strong constraint that often cannot be satisfied perfectly without perfect predictor performance.
• Relationship: Satisfying equalized odds automatically satisfies equal opportunity, as the latter is a relaxation focusing only on the Y=1 condition.

Bias amplification occurs when a model trained on data containing societal biases produces predictions that exhibit those biases to a greater degree than is present in the original training labels. It is a critical risk when generating or using synthetic data.

• Mechanism: The model learns and exacerbates spurious correlations between protected attributes and outcomes.
• Example in Synthesis: If a generative model is trained on historical hiring data biased against a group, the synthetic data it produces may over-represent that bias, leading to worse fairness metrics like statistical parity in downstream models.
• Validation Need: Metrics like statistical parity are used to detect and quantify such amplification in model outputs derived from synthetic datasets.

Disparate impact is a legal doctrine and quantitative measure used to identify unintentional discrimination. It is closely related to statistical parity and is often assessed using the 80% rule (or four-fifths rule).

• The 80% Rule: A selection process has disparate impact if the selection rate for a protected group is less than 80% of the rate for the most favored group.
• Calculation: (Selection Rate_Protected_Group / Selection Rate_Most_Favored_Group) < 0.8.
• Legal Context: In U.S. employment law (Title VII), a finding of disparate impact shifts the burden of proof to the employer to demonstrate job-related necessity.
• Technical Link: A model failing the 80% rule is violating statistical parity. Thus, statistical parity analysis is a core technical method for disparate impact testing.

Adversarial validation is a technique for detecting distribution shift between two datasets—such as real and synthetic data—by training a classifier to distinguish between them. Its failure is a proxy for fidelity.

• Process: 1) Label real data as 0 and synthetic as 1. 2) Train a discriminative model (e.g., a gradient boosting machine) to classify the source. 3) Evaluate its AUC-ROC.
• Interpretation: An AUC near 0.5 (random guessing) suggests the datasets are statistically indistinguishable, indicating high synthetic data quality.
• Connection to Fairness: This method can be adapted to audit for fairness. A domain classifier trained to predict a protected attribute from the model's predictions (or features) should also fail if statistical parity or independence is achieved.